patearn — 14-Pattern Fundamental Quality Methodology — Canonical Reference

One-line definition: patearn is Patearn's rule-based, point-in-time 14-pattern fundamental-quality score for Indian equities — the pure-Python Stage-1 selection lens that reads capital efficiency, operating leverage, balance-sheet strength, valuation asymmetry and ten further quality patterns, gates out governance/leverage blow-ups via five hard disqualifiers, and surfaces the handful of names worth a human deep-dive.

1. What it is

patearn is the project's core analytical lens and namesake — the methodology after which the whole platform is named. It is a disciplined process for identifying Indian **mid-cap multi-baggers before institutional re-rating**, built explicitly to override the intuitive shortcuts that cause most investment mistakes: buying a loud narrative, rationalising away a governance red flag, or confusing price momentum with a live thesis.

Operationally it is two things working together:

1. A rule-based quantitative score (Phase 3). Every surviving company is scored on 14 patterns × 3 signals each, each signal graded No / Partial / Yes, weighted, and rolled into a single Normalised Score (NS, 0–100) with a mandatory sensitivity band, a Pattern Activation Count (PAC), a Quality-Gate floor on the fundamental core, and a tier (T1 → T4, or DISQUALIFIED). This is what the scoring code implements — pure Python, no LLM (Guardrail #4 /). 2. A qualitative process wrapper (the 6 phases). Universe construction → hard filter → quantitative score → qualitative deep dive → entry & sizing → monitoring/exit. The score surfaces candidates; human judgment in claude.ai (Phase 4) verifies whether the story is real.

The governing empirical finding (**) is the one everyone must internalise: the patearn score is a RISK FILTER, not a return ranker.** In the survivorship-aware point-in-time backtest, blow-up rate (>50% loss / 12m) falls monotonically as tier improves, while the long-short return of the score itself is ≈ 0. patearn tells you what not to own and what is high-quality enough to own — the return engine is momentum/accumulation (see the momentum riskadj notes).

2. Our variation vs. the standard technique

Generic fundamental screening (a Piotroski F-Score, a Magic-Formula rank, a "high-ROCE-low-PE" filter) is a static, single-snapshot, count-of-good-things exercise. patearn deliberately departs on five axes, and the departures are the proprietary part:

• 14 hand-designed patterns, not a generic factor bundle. Each pattern encodes a specific thesis about how an Indian mid-cap re-rates (operating leverage as the re-rating engine; a named, funded, time-bounded policy tailwind rather than a vague "sector story"; export inflection into premium markets; institutional neglect as the opportunity window; promoter accumulation while institutions capitulate). The 3-signal / Yes-Partial-No rubric and per-pattern weights are defined in the patterns notes.
• Point-in-time (PIT) rigor. Trends beat snapshots — rising ROCE 12→16→22% is a stronger signal than a static 22%. The PIT reader (the fundamentals asof code) reconstructs exactly what was knowable at each historical date (report-lag-safe), so the score can be backtested with no look-ahead (****).
• Value in rupees, never share count. Every cross-time metric is measured in rupees (capital employed, operating profit, PAT), and dilution is measured directly (PAT-CAGR − EPS-CAGR, or equity-capital CAGR) rather than normalised away — this eliminates a whole class of corporate-action-adjustment bugs (Guardrail #5; realised in the "C" lens, §7).
• Hard disqualifiers are binary, not soft warnings. Five conditions (governance/leverage) disqualify a name regardless of how good the rest of the score is. The framework exists precisely to make "but the tailwind is so strong" un-actionable in real time.
• **The multi-bagger-before-re-rating thesis.** The universe is bounded to a mid-cap band (roughly ₹500 Cr–₹15,000 Cr) — small enough that the re-rating has not already happened, large enough to be liquid and data-covered. A name that has already run >200% caps at a low tier: you'd be buying the story, not the setup.

The calibration benchmark for a Tier-1 setup is Apar Industries FY2021 (≈₹350, pre-~34× move), scored from public disclosures only — documented in the SKILL notes § Calibration Standard.

3. How it works (methodology)

The full mechanism lives in the SKILL notes (process) and the patterns notes (per-signal criteria). Conceptual summary only here — exact weights, thresholds, and the normalisation math are IP and are NOT restated on this page.

The 6-phase process (mandatory sequence, no phase skipped):

**The 14 patterns (conceptual grouping; math in the patterns notes):**

• Fundamental core (the top-5 Quality Gate): ROCE/ROE trajectory· operating leverage (EPS growing faster than revenue)· structural sectoral tailwind (named + funded + time-bounded)· valuation asymmetry· balance-sheet deleveraging. These carry the highest weights, and a name that fails a Quality-Gate floor on this group cannot reach the top tier no matter how good its technical or ownership signals look — this stops "a perfect chart with mediocre fundamentals" from masquerading as a multi-bagger setup.
• Business-quality patterns: product/revenue-mix premiumisation· capex-cycle completion (FCF now being generated)· promoter/insider accumulation· export/global-revenue expansion· institutional neglect (FII at multi-year low, thin analyst coverage).
• Technical / stage patterns (⚠ chart-check required, scored Estimated unless real ATR/RS/volume data is pulled): tight base / VCP breakout· relative strength & stage-of-move.
• Discipline patterns: working-capital efficiency· management quality & disclosure.

Scoring architecture (mechanisms named, constants withheld): each signal is graded No/Partial/Yes; signals not sourced from a citable document are tagged unverified and discounted (they count, but less); the weighted sum is normalised to a 0–100 NS; a mandatory pessimistic / base / optimistic sensitivity band is always reported; PAC counts how many of the 14 patterns are active (breadth check); the Quality Gate enforces a floor on the fundamental core; and names are placed into tiers T1 (strong) → T2 → T3 → T4 (watch), or DISQUALIFIED. A name that has already run >200% in 24 months is tier-capped. The exact weights, thresholds, MAX score, unverified-discount factor and tier cutoffs are in the scoring code + the calculations and weights notes.

The 5 hard disqualifiers (binary, non-negotiable — surfaced, not hidden): promoter pledge over the limit· net D/E too high and still rising· CFO negative for 2+ consecutive years· auditor resignation/qualification/unexplained change· related-party transactions too large a share of revenue. Any one fires → DISQUALIFIED, logged, no entry. Rationale and the six Indian mid-cap failure case studies (Vakrangee, Manpasand, Suzlon, Aban, PC Jeweller, Yes Bank) that motivate them: the failures notes.

What the code can and cannot see: the scoring code runs the quantitative Phase-3 score over numeric fundamentals. The patterns that need narrative/segment/technical inputs (tailwind, export mix, VCP, volume) it scores conservatively as Partial-Estimated and hands off to Phase 4 in claude.ai. Where a real point-in-time trend key is available (ROCE trend, OPM trend, interest coverage, profit acceleration, debtor days, cash-conversion cycle, promoter-rising —), the corresponding signals switch from proxy to verified; the live Telegram snapshot path (no trend keys) is byte-for-byte unchanged (regression-safe).

⚠ Proxy fidelity (Codex -F3, converged). The top-5 Quality-Gate patterns are computed via numeric proxies, not the exact multi-year rules in the patterns notes: e.g. Pattern 1 credits endpoint / 3Y-avg ROCE rising (not three consecutive annual rises or sector-relative improvement); Pattern 2 uses profit-vs-sales growth + OPM level/trend (not EPS CAGR + EBITDA higher-highs + PAT-margin expansion); Pattern 5 uses D/E level + interest-coverage level (not D/E falling + coverage >5x and rising + FCF positive 2-of-3y). The proxies are directionally sound and deliberately conservative, but they are proxies; read a Quality-Gate pass as "broadly clears the fundamental core", not "satisfies every exact criterion". Implementing the exact rules is a queued enhancement.

4. Status, validation & honesty fence

• Deployed and running (Telegram /pt14, nightly batch scoring, web dossier + screener). See §5.
• PIT-backtestable, and backtested — but not validated as standalone alpha. The honest read (survivorship-aware, overlapping monthly observations, descriptive — not significance-tested): the score is a risk filter. Blow-up rate falls monotonically by tier (T2 → T3 → T4 → DISQ); NS long-short ≈ 0. Returns came from accumulation/RS (top-tercile delivery-drift ≈ 11% median 12m vs ~5% weak), and the sweet spot was strong accumulation × good patearn quality (≈11% median / 60% hit / ~2% blow-up).
• T1 never forms in the pure-quant path — NS caps in the low-60s because the tailwind/export/VCP/volume patterns stay Estimated. A T1-capable machine score would require folding the technical-confirmation patterns into the scorer; today T1 is a claude.ai Phase-4 judgment, not a machine output.
• The asset is the lens, not a return edge. Per the patearn-charter: the durable product is PIT rigor + under-covered primary data + the analytical selection lens — an evidence machine — not a backtested alpha strategy. Do not claim a validated return edge for the score itself. It ranks/compares on reproducible, measurable inputs only; interpretive content informs but never ranks.
• "C" (capital-allocation) consumption: the modern quality composite (§7) is used as a blend/tilt and a percentile lens — it subsumes the older ad-hoc 4-metric quality read but is never a hard veto or a standalone ranker ( keeps the C/A/B data lenses veto-/filter-only). Where C is actually consumed inside a return construct (the quality × momentum blend) is documented in the momentum riskadj notes.

5. Where it lives (code· routes· DB· timers)

Code

• the scoring code — the rule-based 14-pattern scorer (Phase 3). score_fundamentals , hard-disqualifier check, sensitivity band, QG, tier, Telegram formatter.
• the capital allocation code — the "C" capital-allocation quality composite (ROIIC, ROCE level+trend, dilution drag, debt-funding share, growth efficiency), PIT-safe, with a separate model='financial' path for lenders.
• the fundamentals asof code — point-in-time fundamentals reader (research.db → scorer dict; no look-ahead; provides the real trend keys).
• the score batch code — bounded, prioritized batch scorer for surfaced names only (honors — never a bulk 5,000-stock scrape).
• the news feed code — twice-daily poller; EARNINGS items trigger the Stage-1 rule-based score.
• the patearn code — the patearn system prompt + Haiku/Sonnet wrapper for the /analyze deep path.

Telegram commands

• /pt14 TICKER — rule-based 14-pattern score (₹0, no LLM).
• /analyze TICKER — full Haiku patearn analysis (use sparingly, ~₹2).
• /conviction — cross-pillar shortlist; pt14 quality shown as a ★ confirmation.

Web routes (/dash)

• Per-stock pt14 card / dossier on the stock page (reads cached pattern_scores).
• Screener column-groups + the Conviction shortlist (/dash/conviction) — pt14 is surfaced as confirmation, not a gate (LEFT JOIN; a ★ marks quality-confirmed, ✗ flags hard-disqualified).

DB tables (the db code)

• pattern_scores — id, symbol, scored_at, pws, ns_base/pessimistic/optimistic, pac, tier, qg_pass, hard_disqualified, disqualifier_reasons, detail_json.
• capital_allocation_scores — ca_score, ca_pctile, ca_tier, model, roiic, roce_latest/avg/trend, roe_latest, roa_latest, dilution_drag, debt_funding_share, growth_efficiency, coverage, detail_json. UNIQUE(symbol, as_of); module-owned schema; PIT-safe.
• fundamentals / fundamentals_history — the inputs (see §6).

Timers (systemd on the VPS)

• hermes-pt14batch.timer — daily bounded batch pt14 scoring of surfaced names, after the signals chain.
• capital_allocation batch — python -m src.automation.capital_allocation --backfill (validated live 2026-07-01, 1,900 names; not yet nightly-wired — see §9).
• Nightly XBRL fundamentals ingest + shareholding-pattern XBRL timer (16:45) feed the inputs.

6. Data & provenance

Bhav copy + delivery = primary source. NSE bhav copy (prices, volumes, delivery), the rolling delivery-value-per-trade signals, and the point-in-time price chain are all authentic NSE data. No concern here.

Fundamentals = the KNOWN primary-sources exception, under active remediation. Disclose this wherever the score is shown.

• patearn's fundamentals historically came via the screener code → Screener.in. This is the copyright-risk exception flagged by Guardrail #8 (primary sources only: NSE/BSE/SEBI/XBRL). Do not extend the Screener dependency.
• The standing remediation (session 73): fundamentals move to primary-source NSE-XBRL forward ingest — the same financial-results XBRL documents the regulator receives — with a per-symbol series-continuity gate (a symbol is only appended once its recent XBRL filings reconcile with the historical series on the rupee core, so restated/netted definitions can't corrupt cross-boundary CAGRs). Every row carries a source tag (NSE-XBRL-CONSO/-SA; NULL = the frozen Screener era). (session 75) extended this to banks/lenders via tag-detection mapping; the shareholding-pattern XBRL feed adds PIT promoter/FII/DII/pledge.
• The legacy Screener series is FROZEN — kept for history, never re-scraped for new coverage. New coverage arrives only through the XBRL path.
• Remaining: Phase 3 = historical backfill, then delete the screener code entirely. Full design + reconciliation evidence: the fundamentals xbrl migration notes.
• The 24-year historical archive (research.db.fundamentals_history, ~1,983 syms × 2002–2026) that powers the PIT backtest and the "C" lens is treated as a frozen, PIT-safe base —.

7. Terminology canon

• patearn — the methodology and its 14-pattern quality score (this page). The project's namesake analytical lens. Lower-case; a coined term (about patterns of earnings/quality).
• The 14 patterns — the hand-designed pattern set (ROCE trajectory … relative strength); definitions in the patterns notes. Grouped into a fundamental core (the Quality-Gate five), business-quality, technical, and discipline patterns.
• NS (Normalised Score) — the 0–100 rolled-up patearn score, always reported with its pessimistic/base/optimistic band.
• Quality Gate (QG) — the floor a name must clear on the top-5 fundamental patterns to be tier-1-eligible.
• PAC (Pattern Activation Count) — how many of the 14 patterns are active; a breadth/robustness check alongside NS.
• Tier — T1 (strong) / T2 / T3 / T4 (watch) / DISQUALIFIED.
• Hard disqualifier — one of the five binary governance/leverage conditions that voids a name outright (≠ a soft "concern").
• Stage-1 screen — the pure-Python, rule-based quantitative pass (Phases 1–3). NOT LLM-driven (/Guardrail #4). Distinct from Phase 4, the qualitative claude.ai deep dive.
• "C" (capital allocation) — the modern derived quality composite (the capital allocation code): has management compounded incremental rupee capital well (ROIIC, ROCE level+trend, dilution, debt-funding, growth efficiency), or just grown size? Consumed as a blend/tilt, never a standalone ranker.
• ⚠ Disambiguation — the patearn SCORE vs. the momentum ENGINE. patearn is the quality/risk-filter selection lens (this page). It is not the return-generating construct. The momentum / risk-adjusted engine (where returns and the quality × momentum blend live) is a separate lens — see the momentum riskadj notes. Keeping these two straight is load-bearing: established that patearn filters and momentum ranks.

8. Decision & session history

• / — Stage-1 screening is rule-based Python, not LLM (cost + precision + reproducibility); do not pre-scrape Screener for thousands of stocks — wait for results events, cache, grow incrementally.
• ** — Phase-4 deep analysis happens in claude.ai under the $20/mo subscription, not the API** (deliberate cost control).
• / Doctrine D — sector-adapted patearn for financials (HFC/NBFC/bank): ROCE→ROE/ROA, D/E gate replaced by GNPA/CAR/ALM, etc.; without adaptation lenders score misleadingly low. The "C" lens realises this with a data-intrinsic model='financial' path.
• / — the Conviction cross-pillar shortlist (RS leader × accumulation × entry read, pt14 as ★ confirmation) + bounded batch pt14 scoring of surfaced names (lights up the Quality pillar without violating).
• ** — rank/compare on measurable, reproducible** inputs only; interpretive content informs but never ranks.
• (session 37, 2026-06-23) — the pivotal decision. PIT fundamentals → backtestable patearn score; the score is a RISK FILTER, returns come from accumulation. Blow-up rate falls by tier; NS long-short ≈ 0; sweet spot = accumulation × quality.
• (session 63, 2026-07-02) — data-roadmap ROI: build C (capital-allocation) → A (insider/promoter) → B (credit ratings); C is a free derived layer over data already held. Validated live 2026-07-01 (Nestlé ~100, TCS 88 vs YesBank 21, JPPower 28); financials carve-out 2026-07-02.
• (session 73) / (session 75) — fundamentals go primary-source NSE-XBRL with a series-continuity gate; Screener history frozen; bank mapper added. The Guardrail-#8 remediation.

9. Open items / frozen work

• XBRL migration, Phase 3 (the big one). Historical backfill of NSE-XBRL fundamentals, then delete the screener code. Until then the Screener input remains the disclosed exception. Owner doc: the fundamentals xbrl migration notes.
• Nightly-wire the "C" batch. capital_allocation --backfill is validated live but still run manually; wire it into the nightly chain ( wired the C/A/B data timers; the score batch itself is the residual).
• T1-capable machine score. Fold the technical-confirmation patterns (VCP, volume — already in ml_panel/bhav) into the scorer so a machine NS can actually reach T1 ( identified this as the unlock).
• C-blend re-check. Re-validate the quality × momentum blend/tilt (does layering C on the momentum sleeve add alpha or cut drawdown?) — tracked in the momentum riskadj notes; QUAL_MOM hints it can, unproven.
• ✅ Reconciled — code↔methodology mapping now explicit. the patterns notes and the scoring code agree exactly on the top-5 Quality Gate; their patterns 6–14 are a deliberate reorganization to what is computable from fundamentals, so the numbers/labels don't line up 1:1 (code#6 Promoter = methodology#8; code#8 Institutional Neglect = methodology#10; code#9/#10 Earnings Momentum/Margin Expansion are computable adjuncts; methodology#7/#13/#14 are Phase-4-only). The authoritative map now lives in the patterns notes § "Implementation mapping (the scoring code)", and the scoring code's stale # Pattern weights from the patterns notes comment is corrected with per-line methodology annotations. Documentation only — scoring output is byte-identical (the desk's call: document the mapping, don't change behavior). Actually re-numbering/re-weighting the live scorer to the canonical methodology is deliberately deferred — it moves every NS + touches pattern_scores.detail_json + the surfaces + AUD-15, so it stays a separate signed decision.

10. Sources of truth

Methodology (resources — two levels up from the strategies folder):

• the SKILL notes — the 6-phase process, scoring architecture, calibration standard.
• the patterns notes — the 14 patterns, per-signal Yes/Partial/No criteria (the pattern math).
• the failures notes — 6 failure case studies + the pre-entry anti-signal checklist.
• the exit protocol — the 4-mode Hold/Trim/Re-evaluate/Exit decision tree.

Product / doctrine:

• the patearn charter — operating doctrine (evidence machine > signal shop; trust is the product).
• the strategy ledger — Tier-3 lens row: "patearn· 14-pattern PIT fundamental quality score· deployed; PIT-backtestable but not run as alpha."
• the calculations and weights notes — the ONE canonical explainer for the numbers/weights (never restated elsewhere).
• the fundamentals xbrl migration notes — the Screener → NSE-XBRL remediation.

Companion strategy lens:

• the momentum riskadj notes — the momentum / risk-adjusted return engine and where "C" is consumed (the quality × momentum blend). patearn filters; momentum ranks.

the project's running record sections: § Decision log (///////**/**///)· § Database schema (pattern_scores, capital_allocation_scores, fundamentals)· § Key file paths· § Telegram bot commands· § Session log (Sessions 37, 63, 73, 75).

Memory (user auto-memory, by slug): primary-intent-north-star (patearn = the best analytical tool for Indian equity research; never leak formulas)· fundamentals-archive-built (the 24y PIT archive; do not re-collect)· dataset-roadmap-c-a-b (C/A/B lenses stay veto-only)· predictive-attributes-finding (momentum the only surviving factor; C/A/B veto-only)· data-sourcing-primary-only (primary sources; never Screener/vendors).